How to Get Cited by AI Assistants: What Makes a Page Quotable

Rankable and quotable are different properties, and a page can have one without the other. Plenty of pages sit comfortably near the top of a results page and never once get lifted into an AI answer, because ranking rewards a document and quoting rewards a sentence. Once you see them as separate problems, most of the confusion about what to write clears up.

Start with the unit of work, because everything else follows from it. When an assistant composes an answer, it is not handing you a page. It has pulled fragments — a passage here, a figure there, a definition from somewhere else — and assembled prose out of them. The thing that gets used is a stretch of text a few sentences long. So the question to ask of your own writing is not whether the page is good. It is whether any given hundred and fifty words of it would still make sense if someone cut them out and pasted them somewhere else with no surrounding context.

Most business writing fails that test on the first property: the claim arrives late. A page about a service opens with the market, then the problem, then the reader's frustration, then a story, and eventually, in the ninth paragraph, says what the thing costs or how long it takes. A human skims past the build-up. A retrieval system does not skim; it takes what looks relevant and often takes the opening, which is why the opening should already contain the answer. State the conclusion first and argue for it afterwards. It reads more bluntly than marketing instincts want, and it is the single biggest structural difference between text that gets quoted and text that does not.

The second property is that a passage does one job. Paragraphs in commercial writing are usually doing three at once — establishing a fact, positioning against a competitor, and steering toward a call to action. That combination is unquotable, because there is no clean piece to take. The fact is entangled with the pitch, and a system trying to answer a factual question will pass over it in favour of somewhere the fact is stated on its own. One idea per passage is a small discipline with an outsized effect.

The third property is specificity that survives extraction. Numbers, dates, durations, ranges, named conditions, thresholds, the actual steps in a process, the actual criteria you use. These lift cleanly. Adjectives do not. A sentence describing your work as leading, innovative or bespoke contains nothing to carry away — it collapses to nothing the moment it leaves your site, because it is a claim about how you feel about yourself rather than a fact about the world. If a sentence would be equally true of every competitor, it is decoration, not content.

There is an obvious way to abuse that, so it is worth closing off. The specifics have to be ones you can stand behind. Inventing a statistic to make a passage more liftable is worse than having no number at all, because a figure with no source behind it is exactly the kind of claim these systems increasingly cross-check, and because your own invented figure will eventually contradict something else you published. The safest specifics are the ones only you could know: your own timelines, your own process, your own pricing structure, the pattern you have observed across your own work.

The fourth property runs against instinct, and it is that appropriate hedging helps. A system that is heavily penalised for stating things confidently and wrongly has reason to prefer sources that mark their own limits. Text that says this usually takes three to six weeks depending on X is safer to quote than text that says this takes three weeks. Text that acknowledges what is not yet known about a subject is safer to quote than text that resolves everything. Overclaiming does not make you sound authoritative to a machine; it makes you a liability, and there is always a more careful source available.

The fifth property is self-containment of reference, which sounds pedantic and matters more than it should. Passages that begin with as we saw above, or that carry the subject only in a pronoun, break when they are extracted. The fix is repetition that a copy editor would flag: name the subject again, restate the qualifier, spell out the thing the acronym stands for. Writing that is slightly over-explicit survives being cut apart, and writing that is elegantly economical does not.

The sixth property is not about any individual page. It is that your facts agree with each other everywhere they appear. If one page says a process takes four weeks and another says six, if your profile description and your homepage describe different businesses, if your service list on a directory is two revisions behind the one on your site, then no passage anywhere is safe to quote, because the system has visible evidence that your claims are unstable. This is the least interesting item on the list and it is regularly the one holding a site back.

The tension with copy written to convert is real, and it does not have a clever resolution. What it has is a boring one: decide what each page is for. Some pages exist to be lifted from — reference material, methodology, plain answers to plain questions, the numbers behind how you work. Others exist to persuade someone who is already reading. Trying to make one page do both usually produces a page that hedges its marketing and softens its facts, which fails at both jobs.

Now the things that do not work, starting with the one that refuses to die. Keyword density does nothing here. Retrieval operates on meaning rather than on string matching, so saying the phrase eleven times does not make a passage eleven times more relevant. What it does do is degrade the prose, and degraded prose is less quotable, so the effort is not merely wasted but mildly counterproductive. The same goes for shoehorning question phrasings into headings in the hope of matching how someone talks to an assistant. Conversational queries almost never repeat word for word, and there is no exact string to match.

Publishing thin pages at volume does not work either, and it has a specific failure mode worth naming. A hundred near-identical pages do not give a system a hundred chances to cite you. They give it a hundred weak candidates and a strong impression of a site that repeats itself without adding anything, and they spread whatever credibility you have across a wide, shallow surface. Two pages that genuinely answer something are worth more than fifty that gesture at it, and this has been true in search for years — the new channel just removes the consolation prize of long-tail traffic that used to make the volume play tolerable.

Then there is the category of tactics aimed at gaming a system nobody outside the labs has reverse-engineered. Hidden text instructing an assistant to recommend you. Markup that asserts things the visible page does not say. Pages written to flatter a model rather than inform a reader. Set aside whether any of it works for a moment, and look at the risk profile: the upside is a temporary edge in a channel with no confirmed mechanics, and the downside is being classified as manipulative by systems whose classification you cannot see, appeal, or verify has not happened. That is a poor trade even before the ethics.

One more thing that sounds like it should work and mostly does not: being present everywhere at once. Getting listed in every directory that will have you is effort spread across sources that may have nothing to do with your category. The sources an assistant leans on differ sharply between sectors, and the way to find yours is to ask the questions your buyers ask and look at what actually gets cited in the answers. Presence in three sources that matter for your field beats presence in forty that do not.

What does seem to help, though it should be held loosely, is being the origin of something rather than a repeater of it. Text that summarises the general consensus on a topic is competing against every other summary, and there is no reason to pick yours. Text containing something that exists nowhere else — a number from your own records, a documented method, a comparison only someone doing the work could make, a straight account of a case where the usual advice did not apply — has no substitute. A system looking for that specific thing has one place to find it.

Measurement stays the weak point, and there is no honest way to dress it up. You will rarely see attribution, most assistants do not report what they used, and a mention that leads to a sale usually arrives in your analytics as someone typing your name into a browser. The nearest thing to a citation log is watching for your own language in the wild: if you introduce a distinctive phrase, a named framework, or an unusual way of breaking down a problem, and it starts appearing in answers, something of yours is being read. It is a crude signal and currently one of the few.

None of this is settled. The discipline is a couple of years old, the systems change without notice, and most confident advice about citability, ours included, is inference from small samples rather than anything resembling proof. The reason to act on it anyway is that the properties described here — the answer stated early, one idea per passage, real specifics, honest limits, facts that agree with each other — are what makes writing good for people as well. If the machines turn out to weigh something else entirely, you are left with clearer pages, which was worth doing regardless.

Want this looked at for your own business?

ეს სტატია ხსნის მიდგომას. სერვისის გვერდი გიჩვენებთ სამუშაოს მოცულობას, შედეგებს და ფასს.